AI and Deep Learning in Biometric Security: Trends, Potential, and Challenges 2020032531, 2020032532, 9780367422448, 9781003003489, 9780367672515

Table of contents :
Cover
Half Title
Series Page
Title Page
Copyright Page
Table of Contents
Preface
Editors
Contributors
Chapter 1 Deep Learning-Based Hyperspectral Multimodal Biometric Authentication System Using Palmprint and Dorsal Hand Vein
1.1 Introduction
1.2 Device Design
1.3 System Implementation
1.3.1 ROI Extraction
1.3.1.1 Hyperspectral Palmprint ROI Extraction
1.3.1.2 Hyperspectral Dorsal Hand Vein ROI Extraction
1.3.2 Feature Extraction
1.3.3 Feature Fusion and Matching
1.4 Experiments
1.4.1 Multimodal Hyperspectral Palmprint and Dorsal Hand Vein Dataset
1.4.2 Optimal Pattern and Band Selection
1.4.3 Multimodal Identification
1.4.4 Multimodal Verification
1.4.5 Computational Complexity Analysis
1.5 Conclusions
Acknowledgements
References
Chapter 2 Cancelable Biometrics for Template Protection: Future Directives with Deep Learning
2.1 Introduction
2.2 Template Protection
2.2.1 Consequences of Template Compromise
2.2.2 Template Protection Techniques
2.2.3 Comparative Analysis between Template Protection Techniques
2.2.4 Fundamental Requirements of Template Protection Techniques
2.2.5 Potential Attacks on Template Protection Techniques
2.3 Role of Deep Learning Approaches in Biometrics
2.3.1 Deep Learning in Face Recognition
2.3.2 Deep Learning in Iris Recognition
2.3.3 Deep Learning in Fingerprint Recognition
2.3.4 Deep Learning in Other Biometric Traits
2.4 Related Work: Template Protection
2.4.1 Biometric Encryption
2.4.2 Biometric Cryptosystems
2.4.3 Cancelable Biometrics
2.4.3.1 Deep Learning-Based Cancelable Techniques
2.4.3.2 Deep Learning versus Non-deep Learning Cancelable Techniques
2.5 Performance Measures and Datasets in Cancelable Biometrics
2.5.1 Performance Measures for Non-invertibility Analysis
2.5.2 Performance Measures for Unlinkability Analysis
2.5.3 Performance Measures for System Usability Analysis
2.5.4 Performance Measures for Revocability Analysis
2.5.5 Databases Used in Cancelable Biometrics
2.6 Comparative Performance Analysis: Cancelable Biometrics
2.7 Conclusions and Future Prospective of Deep Learning in Biometrics
References
Chapter 3 On Training Generative Adversarial Network for Enhancement of Latent Fingerprints
3.1 Introduction
3.2 Related Work
3.3 Proposed Algorithm
3.3.1 Problem Formulation and Objective Function
3.3.2 Training Data Preparation
3.3.3 Network Architecture and Training Details
3.4 Performance Evaluation
3.4.1 Databases and Tools Used
3.4.2 Evaluation Criteria
3.5 Results and Analysis
3.6 Challenges Observed
3.7 Conclusions
Acknowledgements
References
Chapter 4 DeepFake Face Video Detection Using Hybrid Deep Residual Networks and LSTM Architecture
4.1 Introduction
4.2 Related Work
4.2.1 Categories of Face Manipulations
4.2.1.1 Face Synthesis
4.2.1.2 Face Swap
4.2.1.3 Facial Attributes
4.2.1.4 Face Expression
4.2.2 DeepFakes Detection
4.3 Proposed DeepFake Videos Detection Framework
4.3.1 Convolutional Neural Networks (CNNs)
4.3.2 Long Short-Term Memory (LSTM)
4.3.3 Residual Neural Network (ResNet)
4.4 Experiments
4.4.1 Datasets
4.4.1.1 DeepFakeTIMIT Dataset
4.4.1.2 Celeb-DF Dataset
4.4.2 Figures of Merit
4.4.3 Experimental Protocol
4.4.4 Experimental Results
4.5 Challenges and Future Research Directions
4.6 Conclusions
Notes
References
Chapter 5 Multi-spectral Short-Wave Infrared Sensors and Convolutional Neural Networks for Biometric Presentation Attack Detection
5.1 Introduction
5.2 Definitions
5.3 Related Works
5.4 Proposed PAD Method
5.4.1 Hardware: Multi-Spectral SWIR Sensor
5.4.2 Software: Multi-Spectral Convolutional Neural Networks
5.4.2.1 Multi-Spectral Samples Pre-Processing
5.4.2.2 CNN Models
5.4.2.3 Score Level Fusion
5.5 Experimental Setup
5.5.1 Database
5.5.2 Evaluation Metrics
5.5.3 Experimental Protocol
5.6 Experimental Evaluation
5.6.1 Baseline: Handcrafted RGB Conversion
5.6.2 Input Pre-Processing Optimisation
5.6.3 Final Fused System
5.7 Conclusions and Future Research
Acknowledgements
References
Chapter 6 AI-Based Approach for Person Identification Using ECG Biometric
6.1 Introduction
6.2 ECG and Related Work
6.2.1 Advantages of ECG Biometric
6.2.2 Literature Review
6.3 Methodology Adopted
6.3.1 Feature Extraction
6.4 Classifier
6.4.1 Artificial Neural Network (ANN)
6.4.2 Support Vector Machine (SVM)
6.5 Experiments and Results
6.6 Conclusions
References
Chapter 7 Cancelable Biometric Systems from Research to Reality: The Road Less Travelled
7.1 Introduction
7.2 Cancelable Biometric Systems: Introduction and Review
7.2.1 Conventional Template Transformation Techniques
7.2.2 Role of Deep Learning in Biometrics and Need for Privacy
7.2.3 Neutral Network-Based Template Transformation Techniques
7.3 Experimental Reporting
7.4 Real-Life Challenges for Applications of Cancelable Biometric Systems
7.5 Conclusions and Foresights
References
Chapter 8 Gender Classification under Eyeglass Occluded Ocular Region: An Extensive Study Using Multi-spectral Imaging
8.1 Introduction
8.1.1 Our Contributions
8.2 Related Works
8.2.1 Visible Spectrum
8.2.2 Near-Infra-Red Spectrum
8.2.3 Visible and Near-Infra-Red Spectrum
8.2.4 Multi-Spectral Imaging
8.3 Database
8.3.1 Data Preprocessing
8.4 Proposed Method
8.4.1 Spectral Bands Selection
8.4.2 Feature Extraction
8.4.3 Classification
8.5 Experiments and Results
8.5.1 Experimental Evaluation Protocol
8.5.2 Evaluation 1: Without-Glass v/s Without-Glass
8.5.2.1 Individual Band Comparison
8.5.2.2 Fused Band Comparison
8.5.3 Evaluation 2: Without-Glass v/s With-Glass
8.5.3.1 Individual Band Comparison
8.5.3.2 Fused Band Comparison
8.6 Conclusions
Acknowledgement
References
Chapter 9 Investigation of the Fingernail Plate for Biometric Authentication using Deep Neural Networks
9.1 Introduction
9.1.1 Motivation and Scope of Present Work
9.2 Related Work
9.3 Sample Acquisition and ROI Extraction
9.3.1 Sample Acquisition
9.3.2 ROI Extraction
9.4 Feature Extraction
9.4.1 Transfer Learning using AlexNet
9.4.2 Transfer Learning using ResNet-18
9.4.3 Transfer Learning using DenseNet-201
9.5 Multimodal System Design
9.5.1 Score-Level Fusion
9.5.2 Rank-Level Fusion
9.5.2.1 Logistic Regression Method
9.5.2.2 Mixed Group Rank
9.5.2.3 Inverse Rank Position
9.5.2.4 Nonlinear Weighted Methods
9.6 Experiments, Results, and Analyses
9.6.1 Performance of Fingernail Plates in Verification Systems
9.6.1.1 Performance of Fingernail Plates in Unimodal Verification Systems
9.6.1.2 Performance of Fingernail Plates in Multimodal Verification Systems
9.6.2 Performance of Fingernail Plates in Identification Systems
9.6.2.1 Performance of Fingernail Plates in Unimodal Identification Systems
9.6.2.2 Performance of Fingernail Plates in Multimodal Identification Systems
9.7 Challenges and Scope of Fingernail Plates in Biometrics
9.8 Conclusions and Future Scope
References
Chapter 10 Fraud Attack Detection in Remote Verification Systems for Non-enrolled Users
10.1 Introduction
10.2 Related Work
10.2.1 Remote Authentication Framework Using Biometrics
10.2.2 Image Manipulation and Deep Learning Techniques
10.3 Fake ID Card Detection for Non-enrolled Users
10.3.1 Databases
10.3.2 Hand-Crafted Feature Extraction (BSIF, uLBP, and HED)
10.3.3 Automatic Feature Extraction (CNN)
10.4 Experiments and Results
10.4.1 Feature Extraction Classification
10.4.2 Classification Using CNN Algorithms
10.4.2.1 Small-VGG Trained from Scratch
10.4.2.2 Pre-trained VGG16 Model and Bottleneck
10.4.2.3 Pre-trained VGG16 Model and Fine-Tuning
10.5 Conclusions
Acknowledgement
References
Chapter 11 Indexing on Biometric Databases
11.1 Introduction
11.2 Indexing Facial Images
11.2.1 Predictive Hash Code
11.2.2 Results
11.3 Indexing Fingerprint Images
11.3.1 Coaxial Gaussian Track Code
11.3.2 Results
11.4 Indexing Finger-Knuckle Print Database
11.4.1 Boosted Geometric Hashing
11.4.2 Results
11.5 Indexing Iris Images
11.5.1 Indexing of Iris Database Based on Local Features
11.5.2 Results
11.6 Indexing Signature Images
11.6.1 KD-Tree-Based Signature Database Indexing
11.6.2 Results
11.7 Conclusion
References
Chapter 12 Iris Segmentation in the Wild Using Encoder-Decoder-Based Deep Learning Techniques
12.1 Introduction
12.2 Deep Learning for Segmentation
12.3 Related Work
12.3.1 Non-Deep Learning-Based Methodologies
12.3.2 Deep Learning-Based Methodologies
12.4 Data Sets and Evaluation Metrics
12.4.1 Data sets
12.4.2 CASIA
12.4.2.1 UBIris v1 and UBIris v2
12.4.2.2 NICE-I and NICE-II
12.4.2.3 ND-Iris-0405
12.4.2.4 IITD
12.4.2.5 CSIP
12.4.2.6 MICHE-I and MICHE-II
12.4.2.7 SBVPI
12.4.2.8 IRISSEG-CC
12.4.2.9 IRISSEG-EP
12.4.2.10 MMU1 and MMU2
12.4.2.11 OpenEDS
12.4.2.12 iBUG
12.4.3 Performance Metrics
12.4.3.1 Jaccard Index (JI)
12.4.3.2 Mean Segmentation Error
12.4.3.3 Nice2 Error
12.5 Experimentation
12.6 Challenges Identified and Further Direction
12.7 Conclusion
Acknowledgements
References
Chapter 13 PPG-Based Biometric Recognition: Opportunities with Machine and Deep Learning
13.1 Introduction
13.2 Photoplethysmogram (PPG)
13.3 Literature Review
13.4 Multi-Feature Approach for PPG Biometric
13.4.1 Signal Acquisition
13.4.2 Baseline Wander and Noise Removal
13.4.3 Feature Extraction
13.4.3.1 Pulse Extraction and Normalisation
13.4.3.2 First- and Second-Order Derivatives
13.4.3.3 Autocorrelation
13.5 Classification
13.6 Experiments and Results
13.7 Conclusions
References
Chapter 14 Current Trends of Machine Learning Techniques in Biometrics and its Applications
14.1 Introduction
14.1.1 Biometric Systems
14.1.2 Brain Stroke
14.1.2.1 Risk Factors
14.1.2.2 Blood Pressure
14.1.2.3 Heart Disease
14.1.2.4 Diabetes Mellitus
14.1.2.5 Cholesterol
14.1.2.6 Smoking
14.1.2.7 Alcohol
14.1.2.8 Other Risk Factors
14.1.3 Face Recognition
14.1.4 Motivation to Machine Learning Techniques
14.2 Related Work
14.2.1 Review on Brain Stroke
14.2.2 Review on Face Recognition
14.2.3 Brain Stroke Prediction System
14.2.3.1 Image Acquisition
14.2.3.2 Pre-processing
14.2.3.3 Feature Extraction
14.2.3.4 Classification Using Machine Leaning Algorithms
14.2.3.5 Construction of Convolutional Neural Network
14.2.4 Face-Recognition System
14.3 Discussion and Results
14.3.1 Performance of Brain Stroke
14.3.2 Performance of Face Recognition
14.4 Future Scope
14.5 Conclusion
References
Index

Citation preview

AI and Deep Learning in Biometric Security

Artificial Intelligence (AI): Elementary to Advanced Practices Series Editors: Vijender Kumar Solanki, Zhongyu (Joan) Lu, and Valentina E. Balas In the emerging smart city technology and industries, the role of artificial intelligence is getting more prominent. This AI book series will aim to cover the latest AI work, which will help the naïve user to get support to solve existing problems, and for the experienced AI practitioners, it will shed light on new avenues in the AI domains. The series will cover the recent work carried out in AI and associated domains and it will also cover a broad scope of application areas such as biometric security, Pattern recognition, NLP, Expert Systems, Machine Learning, Block-Chain, and Big Data. The work domain of AI is quite deep, so it will be covering the latest trends that are evolving with the concepts of AI, and it will be helpful to those who are new to the field, practitioners, students, and researchers to gain some new insights.

Cyber Defense Mechanisms Security, Privacy, and Challenges Gautam Kumar, Dinesh Kumar Saini, and Nguyen Ha Huy Cuong

Artificial Intelligence Trends for Data Analytics Using Machine Learning and Deep Learning Approaches K. Gayathri Devi, Mamata Rath, and Nguyen Thi Dieu Linh

Transforming Management Using Artificial Intelligence Techniques Vikas Garg and Rashmi Agrawal

AI and Deep Learning in Biometric Security Trends, Potential, and Challenges Gaurav Jaswal, Vivek Kanhangad, and Raghavendra Ramachandra For more information on this series, please visit: https://www.crcpress.com/ Artificial-Intelligence-AI-Elementary-to-Advanced-Practices/book-series/ CRCAIEAP

AI and Deep Learning in Biometric Security Trends, Potential, and Challenges

Edited by

Gaurav Jaswal, Vivek Kanhangad, and Raghavendra Ramachandra

MATLAB® and Simulink® are trademarks of The MathWorks, Inc. and are used with permission. The MathWorks does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB® and Simulink® software or related products does not constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the MATLAB® and Simulink® software. First edition published 2021 by CRC Press 6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742 and by CRC Press 2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN © 2021 Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, LLC Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, access www.copyright. com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact mpkbookspermissions@ tandf.co.uk Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging-in-Publication Data Names: Jaswal, Gaurav, editor. | Kanhangad, Vivek, editor. | Ramachandra, Raghavendra, editor. Title: AI and deep learning in biometric security : trends, potential, and challenges / edited by Gaurav Jaswal, Vivek Kanhangad, and Raghavendra Ramachandra. Description: First edition. | Boca Raton, FL : CRC Press, 2021. | Series: Artificial intelligence (AI) : elementary to advanced practices | Includes bibliographical references and index. Identifiers: LCCN 2020032531 (print) | LCCN 2020032532 (ebook) | ISBN 9780367422448 (hardback) | ISBN 9781003003489 (ebook) Subjects: LCSH: Biometric identification. | Artificial intelligence. Classification: LCC TK7882.B56 A53 2021 (print) | LCC TK7882.B56 (ebook) | DDC 006.2/48—dc23 LC record available at https://lccn.loc.gov/2020032531 LC ebook record available at https://lccn.loc.gov/2020032532 ISBN: 978-0-367-42244-8 (hbk) ISBN: 978-0-367-67251-5 (pbk) ISBN: 978-1-003-00348-9 (ebk) Typeset in Times by codeMantra

Contents Preface......................................................................................................................vii Editors........................................................................................................................ix Contributors...............................................................................................................xi Chapter 1 Deep Learning-Based Hyperspectral Multimodal Biometric Authentication System Using Palmprint and Dorsal Hand Vein..........1 Shuping Zhao, Wei Nie, and Bob Zhang Chapter 2 Cancelable Biometrics for Template Protection: Future Directives with Deep Learning........................................................... 23 Avantika Singh, Gaurav Jaswal, and Aditya Nigam Chapter 3 On Training Generative Adversarial Network for Enhancement of Latent Fingerprints.......................................................................... 51 Indu Joshi, Adithya Anand, Sumantra Dutta Roy, and Prem Kumar Kalra Chapter 4 DeepFake Face Video Detection Using Hybrid Deep Residual Networks and LSTM Architecture..................................................... 81 Semih Yavuzkiliç, Zahid Akhtar, Abdulkadir Sengür, and Kamran Siddique Chapter 5 Multi-spectral Short-Wave Infrared Sensors and Convolutional Neural Networks for Biometric Presentation Attack Detection........ 105 Marta Gomez-Barrero, Ruben Tolosana, Jascha Kolberg, and Christoph Busch Chapter 6 AI-Based Approach for Person Identification Using ECG Biometric........................................................................................... 133 Amit Kaul, A.S. Arora, and Sushil Chauhan Chapter 7 Cancelable Biometric Systems from Research to Reality: The Road Less Travelled................................................................... 155 Harkeerat Kaur and Pritee Khanna

v

vi

Contents

Chapter 8 Gender Classification under Eyeglass Occluded Ocular Region: An Extensive Study Using Multi-spectral Imaging.......................... 175 Narayan Vetrekar, Raghavendra Ramachandra, Kiran Raja, and R. S. Gad Chapter 9 Investigation of the Fingernail Plate for Biometric Authentication using Deep Neural Networks....................................205 Surabhi Hom Choudhury, Amioy Kumar, and Shahedul Haque Laskar Chapter 10 Fraud Attack Detection in Remote Verification Systems for Non-enrolled Users........................................................................... 239 Ignacio Viedma, Sebastian Gonzalez, Ricardo Navarro, and Juan Tapia Chapter 11 Indexing on Biometric Databases..................................................... 257 Geetika Arora, Jagdiah C. Joshi, Karunesh K. Gupta, and Kamlesh Tiwari Chapter 12 Iris Segmentation in the Wild Using Encoder-Decoder-Based Deep Learning Techniques............................................................... 283 Shreshth Saini, Divij Gupta, Ranjeet Ranjan Jha, Gaurav Jaswal, and Aditya Nigam Chapter 13 PPG-Based Biometric Recognition: Opportunities with Machine and Deep Learning............................................................. 313 Amit Kaul and Akhil Walia Chapter 14 Current Trends of Machine Learning Techniques in Biometrics and its Applications........................................................................... 333 B. S. Maaya and T. Asha Index....................................................................................................................... 361

Preface With the growth of data and the increasing awareness about the sensitivity of personal information, people have started to treat their privacy more seriously. Biometric systems have now significantly improved person identification and verification, playing an important role in personal, national, and global security. The recently evolved deep neural networks (DNN) learn the hierarchical features at intermediate layers automatically from the data and have shown many inspiring results for biometric applications. With this motivation, the text offers a showcase of cutting-edge research on the use of DNN in face, nail, finger knuckle, iris, ECG, palm print, fingerprint, vein, and medical biometric systems, and hence focuses on two parts: “Biometrics” and “Deep Learning for Biometrics”. This text highlights original case studies to solve real-world problems on biometric security and presents a broad overview of advanced deep learning architectures for learning domain-specific feature representation for biometrics-related tasks. The book aims to provide an in-depth overview of the recent advancements in the domain of biometric security using artificial intelligence (AI) and deep learning techniques, enabling readers to gain a deeper insight into the technological background of this domain. The text acts as a platform for the decision on the use of advanced architectures of convolutional neural networks, generative adversarial networks, autoencoders, recurrent convolutional neural networks, and graph convolution neural networks for various biometric security tasks such as indexing, gender classification, recognition in the wild, spoofing attacks/liveness detection, quality analysis, ROI segmentation, cross-sensor matching, and domain adaptation. In the text, feasibility studies on medical modalities (ECG, EEG, PPG) have been investigated using AI and deep learning. This book also examines the potential and future perspectives of AI and deep learning towards biometric template protection and multi-spectral biometrics. Overall, the reference provides better readability to readers through its chapter organisation and contains fourteen chapters only. This text/reference is an edited volume by prominent academic researchers and industry professionals in the area of AI and biometric security. It will be essential reading for prospective undergraduate/postgraduate students, young researchers, and technology aspirants who are willing to research in the field of AI and biometric security. Gaurav Jaswal Vivek Kanhangad Raghavendra Ramachandra

vii

viii

Preface

MATLAB® is a registered trademark of The MathWorks, Inc. For product information, Please contact: The MathWorks, Inc. 3 Apple Hill Drive Natick, MA 01760-2098 USA Tel: 508-647-7000 Fax: 508-647-7001 E-mail: [email protected] Web: www.mathworks.com

Editors Dr. Gaurav Jaswal is currently working as post-doctoral researcher at Indian Institute of Technology Delhi, India since January 2020. Before this, he served as Project Scientist (Electrical Engineering) at National Agri-Food Biotechnology Institute Mohali, India. He was research associate at School of Computing and Electrical Engineering, Indian Institute of Technology Mandi, India. He received MTech and PhD degrees in Electrical Engineering from National Institute of Technology Hamirpur in 2018. His research interests are in the areas of multimodal biometrics, medical imaging, and deep learning. He regularly reviews papers for various international journals including IEEE Transactions on Information Forensics and Security (TIFS), IEEE Transactions on Biometrics, Behavior, and Identity Science (T-BIOM), IET Biometrics, and Pattern Recognition Letters. Dr. Vivek Kanhangad is currently working as associate professor, Department of Electrical Engineering, Indian Institute of Technology Indore, India since February 2012. Prior to this, he was visiting assistant professor, International Institute of Information Technology Bangalore, India (June 2010–December 2012). He received PhD from the Hong Kong Polytechnic University in 2010. Prior to joining Hong Kong PolyU, he received MTech degree in Electrical Engineering from Indian Institute of Technology Delhi in 2006 and worked for Motorola India Electronics Ltd, Bangalore for a while. His research interests are in the overlapping areas of digital signal and image processing, pattern recognition with a focus on biometrics and biomedical applications. He regularly reviews papers for various international journals including IEEE Transactions on Information Forensics and Security (TIFS), IEEE Transactions on Cybernetics, IEEE Transactions on Human-Machine Systems, and Elsevier journals – Pattern Recognition and Pattern Recognition Letters. Dr. Raghavendra Ramachandra is currently working as a professor in Department of Information Security and ­ Communication Technology (IIK). He is a member of Norwegian Biometrics Laboratory (http://nislab.no/biometrics_lab) at NTNU Gjøvik. He received B.E. (Electronics and Communication) from University of Mysore, India; MTech (Digital Electronics and Advanced Communication Systems) from Visvesvaraya Technological University, India; and PhD (Computer Science with specialisation of Pattern Recognition and Image Processing) from the University of Mysore, India, and Telcom SudParis, France. His research interest includes pattern recognition, image and video analytics, biometrics, human behaviour analysis, video surveillance, health biometrics, and smartphone authentication. ix

Contributors Zahid Akhtar Department of Computer Science University of Memphis Memphis, Tennessee Adithya Anand Indian Institute of Technology Delhi Delhi, India A.S. Arora Department of Electrical & Instrumentation Engineering Sant Longowal Institute of Engineering and Technology Longowal, India Geetika Arora Department of Computer Science and Information Systems Birla Institute of Technology and Science Pilani Pilani, India T. Asha Department of CSE Banglore Institute of Technology Bengaluru, India Christoph Busch da/sec – Biometrics and Internet Security Research Group Hochschule Darmstadt Darmstadt, Germany Sushil Chauhan Department of Electrical Engineering National Institute of Technology Hamirpur Hamirpur, India

Surabhi Hom Choudhury Department of Electronics & Instrumentation Engineering National Institute of Technology Silchar Silchar, India R.S. Gad Department of Electronics Goa University Taleigao-Plateau, India Marta Gomez-Barrero Fakultät Wirtschaft Hochschule Ansbach Ansbach, Germany Sebastian Gonzalez R+D TOC Biometrics Labs Santiago, Chile Divij Gupta Department of Electrical Engineering Indian Institute of Technology Jodhpur Jodhpur, India Karunesh K. Gupta Department of Electrical and Electronics Engineering Birla Institute of Technology and Science Pilani Pilani, India Gaurav Jaswal Department of Electrical Engineering Indian Institute of Technology Delhi Delhi, India xi

xii

Ranjeet Ranjan Jha School of Computing and Electrical Engineering Indian Institute of Technology Mandi Mandi, India Indu Joshi Indian Institute of Technology Delhi Delhi, India Jagdiah C. Joshi Department of Electrical and Electronics Engineering Birla Institute of Technology and Science Pilani Pilani, India Prem Kumar Kalra Indian Institute of Technology Delhi Delhi, India Amit Kaul Department of Electrical Engineering National Institute of Technology Hamirpur Hamirpur, India Harkeerat Kaur Indian Institute of Technology Jammu Pritee Khanna PDPM Indian Institute of Information Technology, Design and Manufacturing, Jabalpur Jabalpur, India Jascha Kolberg da/sec – Biometrics and Internet Security Research Group Hochschule Darmstadt Darmstadt, Germany Amioy Kumar Client Computing Group Intel Corporation Bangalore Bangalore, India

Contributors

Shahedul Haque Laskar Department of Electronics & Instrumentation Engineering National Institute of Technology Silchar Silchar, India B.S. Maaya Department of CSE Banglore Institute of Technology Bengaluru, India Ricardo Navarro R+D TOC Biometrics Labs Santiago, Chile Wei Nie Department of Computer and Information Science University of Macau Macau, China Aditya Nigam School of Computing and Electrical Engineering Indian Institute of Technology Mandi Mandi, India Kiran Raja Norwegian Biometrics Laboratory Norwegian University of Science and Technology (NTNU) Gjøvik, Norway Raghavendra Ramachandra Norwegian Biometrics Laboratory Norwegian University of Science and Technology (NTNU) Gjøvik, Norway Sumantra Dutta Roy Indian Institute of Technology Delhi Delhi, India

xiii

Contributors

Shreshth Saini Department of Electrical Engineering Indian Institute of Technology Jodhpur Jodhpur, India

Narayan Vetrekar Department of Electronics Goa University Taleigao-Plateau, India

Abdulkadir Sengür Department of Electrical and Electronics Engineering Fırat University Elazig, Turkey

Ignacio Viedma R+D TOC Biometrics Labs Santiago, Chile

Kamran Siddique Department of Information and Communication Technology Xiamen University Malaysia Sepang, Malaysia Avantika Singh School of Computing and Electrical Engineering Indian Institute of Technology Mandi Mandi, India Juan Tapia R+D TOC Biometrics Labs Santiago, Chile Kamlesh Tiwari Department of Computer Science and Information Systems Birla Institute of Technology and Science Pilani Pilani, India Ruben Tolosana Biometrics and Data Pattern Analytics – BiDA Lab Universidad Autonoma de Madrid Madrid, Spain

Akhil Walia Department of Electrical Engineering National Institute of Technology Hamirpur Hamirpur, India

Semih Yavuzkiliç Department of Electrical and Electronics Engineering Fırat University Elazig, Turkey

Bob Zhang Department of Computer and Information Science University of Macau Macau, China Shuping Zhao Department of Computer and Information Science University of Macau Macau, China

1

Deep Learning-Based Hyperspectral Multimodal Biometric Authentication System Using Palmprint and Dorsal Hand Vein Shuping Zhao, Wei Nie, and Bob Zhang University of Macau

CONTENTS 1.1 Introduction.......................................................................................................1 1.2 Device Design....................................................................................................5 1.3 System Implementation..................................................................................... 6 1.3.1 ROI Extraction.......................................................................................6 1.3.1.1 Hyperspectral Palmprint ROI Extraction...............................6 1.3.1.2 Hyperspectral Dorsal Hand Vein ROI Extraction..................8 1.3.2 Feature Extraction............................................................................... 10 1.3.3 Feature Fusion and Matching.............................................................. 13 1.4 Experiments..................................................................................................... 13 1.4.1 Multimodal Hyperspectral Palmprint and Dorsal Hand Vein Dataset......14 1.4.2 Optimal Pattern and Band Selection................................................... 14 1.4.3 Multimodal Identification.................................................................... 17 1.4.4 Multimodal Verification...................................................................... 17 1.4.5 Computational Complexity Analysis................................................... 18 1.5 Conclusions...................................................................................................... 19 Acknowledgements................................................................................................... 19 References................................................................................................................. 19

1.1 INTRODUCTION Biometric recognition system has been widely used in the construction of a smart society. Many types of biometric systems, including face, iris, palmprint, palm vein, dorsal hand vein, and fingerprint, currently exist in security authentication. Palmprint 1

2

AI and Deep Learning in Biometric Security

recognition system is a kind of reliable authentication technology, due to the fact that palmprint has stable and rich characteristics, such as textures, local orientation features, and lines. In addition, a palmprint is user-friendly and cannot be easily captured by a hidden camera device without cooperation from the users. However, palmprint images captured using a conventional camera cannot be used in liveness detection. Palm vein is a good remedy for the weakness of palmprint acquired using a near-infrared (NIR) camera. The vein pattern is the vessel network underneath human skin. It can successfully protect against spoofing attacks and impersonation. Similar to palm vein, dorsal hand vein also has stable vein structures that do not change with age. Besides vein networks, some related characteristics to palmprint such as textures and local direction features can also be acquired. Up to now, palmprint and dorsal hand vein-based recognition methods have achieved competitive performances in the literature. Huang et al. [1] put forward a method for robust principal line detection from the palmprint image, even if the image contained long wrinkles. Guo et al. [2] presented a binary palmprint direction encoding schedule for multiple orientation representation. Sun et al. [3] presented a framework to achieve three orthogonal line ordinal codes. Zhao et al. [4] constructed a deep neural network for palmprint feature extraction, where a convolutional neural network (CNN)-stack was constructed for hyperspectral palmprint recognition. Jia et al. presented palmprint-oriented lines in [5]. Khan et al. [6] applied the principle component analysis (PCA) to achieve a low-dimensionality feature in dorsal hand vein recognition. Khan et al. [7] obtained a low-dimensionality feature representation with Cholesky decomposition in dorsal hand vein recognition. Lee et al. [8] encoded multiple orientations using an adaptive two-dimensional (2D) Gabor filter in dorsal hand vein feature extraction. The palmprint and dorsal hand vein recognition is usually carried out by conventional and deep learning-based methods. The conventional methods need to design a filter to extract the corresponding feature, i.e., local direction, local line, principal line, and texture. These hand-crafted algorithms usually require rich prior knowledge based on the specific application scenario. PalmCode [9] encoded palmprint features on a fixed direction by using a Gabor filter. Competitive code [10] extracted the dominant direction feature by using six Gabor filters. Xu et al. [11] encoded a competitive code aiming to achieve the accurate palmprint dominant orientation. Fei et al. [12] detected the apparent direction from the palmprint image. In addition, Huang et al. [13] put forward a centroid-based circular key-point grid (CCKG) pattern in dorsal hand vein recognition, which extracts local features based on keypoints detection. Deep learning-based algorithms require a mass of training data to train the parameters in the deep convolutional neural network (DCNN). Afterwards, the optimal DCNN can be utilised for classification or convolution feature extraction. However, a mass of training data is usually unavailable for a palmprint or dorsal hand vein recognition task. Especially, the transfer learning technology with DCNN supports an approach that a pretrained DCNN can be fine-tuned with a few training samples for classification in a specific application. Zhao et al. [14] proposed a deep discriminative representation method, which extracted palmprint features from deep discriminative convolutional networks (DDCNs). DDCNs contain a pretrained DCNN and a set of lightened CNNs corresponding to the global and local patches

Multimodal Biometric Authentication System

3

segmented from the palmprint image. Wan et al. [15] trained the VGG depth CNN to extract dorsal hand vein features and used the logistic regression for identification. Deep learning-based methods can be widely applied in generic application scenarios. Increasing research studies have moved to the area of hyperspectral imagery technology in the past decades. Contrary to the traditional imagery technology, not only skin texture but also vascular networks are imaged using the designed hyperspectral imagery system with the specific spectrum setup. In the phase of imaging palmprint or dorsal hand combined hyperspectral technology, more discriminative information from the palmprint or dorsal hand image can be captured achieving a high recognition rate. With more than 60 bands covered in hyperspectral palmprint, the threedimensional (3D) feature was extracted through 3D Gabor filters [16]. Due to the redundant data, hyperspectral palmprint authentication improved but not remarkably when every spectral data were considered in the feature extraction phase. Based on band combination, Shen et al. [17] clustered typical bands in hyperspectral palmprint images for authentication, which performed better compared with in Ref. [16], while Guo et al. [18] applied an approach of k means algorithm for representative band selection in hyperspectral palmprint database to improve performance. What’s more, the band clustering method can decrease computation and increase efficiency in hyperspectral biometrics. As is known, dorsal hand vein and palmprint are concentrated in one hand, which makes it more convenient to collect these two different modalities simultaneously. Based on this observation, the combination of hyperspectral palmprint and dorsal hand biometrics is developed to meet a higher security requirement and to guarantee an exceptional recognition performance. In addition, unimodal biometrics recognition based on a single trait easily suffers from spoofing and other attacks as stated in the literature [19,20]. Table 1.1 illustrates the survey of the current multimodal biometric recognition algorithms. First, it is observed from this table that palmprint and dorsal hand vein have been fused before [21]. However, Ref. [21] and the other methods in Table 1.1 used only two single-spectrum images (one for each modality) to improve the recognition performance. Different from the literature in Table 1.1, this work will study and implement the merging hyperspectral palmprint feature into dorsal hand vein feature to develop a novel hyperspectral multimodal biometric authentication system, which is demonstrated by a flow diagram (refer to Figure 1.1). A hyperspectral acquisition device was utilised for collecting hyperspectral palmprint and dorsal hand images. Then, region of interest (ROI) is detected from hyperspectral palmprint, and dorsal hand images resulted in two corresponding ROI cubes. After ROI extraction, the optimal feature pattern, i.e., local binary pattern (LBP) [22], local derivative pattern (LDP) [9], 2D-Gabor [2], and deep convolutional feature (DCF) [23], is selected for the palmprint and dorsal hand vein, correspondingly. In the pattern selection procedure, each image in the ROI cube is extracted and its features are used in recognition. Thus, the pattern and band which can achieve the highest recognition are treated as the optimal pattern and band for hyperspectral palmprint and dorsal hand images. Afterwards, the feature corresponding to the optimal pattern from palmprint on the optimal band and the feature concerning to the optimal pattern from dorsal hand vein on the optimal band are merged as one feature vector. At last, this fused multimodal feature vector is directly used in matching with the 1-NN classifier.

4

AI and Deep Learning in Biometric Security

TABLE 1.1 The Survey of Multimodal Biometric Recognition Algorithms Literature

Algorithms

[19]

Concatenation

[20] [24]

Combined face-plus-ear image Concatenation

[25] [26] [27] [28]

Concatenation Concatenation Concatenation Fusion

[29]

Fusion

[30]

Concatenation

[31] [32]

Weighted concatenation Feature level

[21]

Score level

[33] [34]

Bayesian decision fusion Score level

[35]

Concatenation

Modalities

Features

Palmprint and hand-geometry Face and ear

Line features; hand lengths and widths PCA

Face and hand

PCA, linear discriminant analysis (LDA) and 9-byte features 2D-Gabor PCA Minutia features PCA Discrete cosine transforms

Face and palmprint Fingerprint and face Side face and gait Palmprint and fingerprint Profile face and ear Palmprint and fingerprint Face and ear Iris, face and fingerprint Palmprint and dorsal hand vein Face and ear Finger-vein and finger shape Face and ear

Speeded up robust features (SURF) Bank of 2D-Gabor

Year 2003 2003 2005

2007 2007 2008 2012 2013 2014

PCA Group sparse representationbased classifier (GSRC) Mean and average absolute deviation (AAD) features CNN features CNN features

2015 2016

2017 2018

CNN features

2017

2016

FIGURE 1.1  The flowchart of the designed system-merged hyperspectral palmprint feature with dorsal hand feature.

Multimodal Biometric Authentication System

5

The major contributions in the chapter are briefly introduced as follows:



1. A novel real-time hyperspectral multimodal biometric authentication system is conceived. It captures hyperspectral hand images by the proposed hyperspectral imaging acquisition device under 53 spectrums in the range of 520–1040 nm with intervals of 10 nm. 2. We collected a big multimodal dataset containing hyperspectral palmprint and dorsal hand images using the designed device. More information about this dataset can be found in Section 1.4.1.

The remaining work is organised as follows. In Section 1.2, the designed capture device is introduced. Following this, the designed system is illustrated in Section 1.3, including ROI and feature extraction as well as multimodal fusion and matching. Extensive experiments and analysis are included in Section 1.4, while Section 1.5 concludes the proposed system.

1.2 DEVICE DESIGN The hyperspectral imaging acquisition system consists of two halogen lamps made by Osram, Inc., one charge coupled device (CCD) camera produced by Cooke, Inc., and one liquid crystal tunable filter manufactured by Meadowlark, Inc. The cost of the setup is approximately USD 6,500.00. The prototype of this acquisition system is illustrated in Figure 1.2. The CCD camera is placed in the middle with one halogen lamp on either side. The halogen lamps produce both visible light and NIR with spectra ranging from 520 to 1,040 nm. The light from the two halogen lamps irradiates on the palm or dorsal hand, and then reflects to the camera sensor for capturing images. A tunable filter is settled ahead of the camera lens and allows a single band to pass through its settings. To obtain stable spectral images, 10 nm is set as the spectral distance in the tunable filter. Therefore, this hyperspectral

FIGURE 1.2  Schematic of our designed hyperspectral imaging device.

6

AI and Deep Learning in Biometric Security

…

…

…

…

FIGURE 1.3  Hyperspectral palm (the upper row) and dorsal hand (the lower row) samples.

imagery acquisition system contains 53 bands in the range of 520–1,040 nm with 10 nm intervals. Each volunteer was asked to grasp a prop making a fist when capturing his\her dorsal hand images. Contrary to an open hand, a closed dorsal hand makes the vascular network more visible achieving discriminant feature exaction. For the palmprint, each individual placed his/her hand on a plate with pegs to somewhat fix their hand, while a cutout was made to expose the palm. Examples of hyperspectral palm and dorsal hand images captured using the designed apparatus are shown in Figure 1.3.

1.3 SYSTEM IMPLEMENTATION First, the ROI detection algorithms for hyperspectral palmprint images and dorsal hand vein images are introduced, respectively. Afterwards, several widely used patterns are presented for feature extraction. At last, a feature fusion strategy is proposed for multimodal recognition of hyperspectral palmprint and dorsal hand vein.

1.3.1 ROI Extraction 1.3.1.1 Hyperspectral Palmprint ROI Extraction It is necessary to conduct ROI extraction from the palm image, due to the fact that the location of the ROI will influence the effectiveness of the extracted feature and  the  recognition performance. Here, we adaptively and reliably detect an ROI from the original palm image, which contains rich and stable characteristics. This step also makes the discriminative characteristics of palmprint separable from the background that contains noise and interference information. In this system, we used the hyperspectral palmprint ROI extraction method, which is based on our previous work in Ref. [36] (refer to Figure 1.4): 1. Image Enhancement: A Laplacian operator with eight neighbourhoods [37] is utilised for sharpness improvement of the original palmprint image. Afterwards, the image quality is much enhanced and will be beneficial for further preprocessing in the next steps (refer to Figure 1.4a and b). The ­utilised Laplacian operator is defined as follows:

7

Multimodal Biometric Authentication System

 0  −1   0



−1 5 −1

0 −1 0

   

2. Binarisation: The Niblack [38] algorithm is a binarisation method which adaptively and locally computes the threshold of the image by performing a convolution. We first transform the enhanced palmprint image into greyscale. Then, a 2D median filter is utilised for noise reduction. In Ref. [39], it has been proved that a 2D median filter can achieve a better performance on denoising in the hyperspectral images. Lastly, we obtain the binary palmprint image using the Niblack method [38] (refer to Figure 1.4b and c). 3. Palm Detection: Given the binarisation palmprint image (refer to Figure 1.4c), we initially locate the tips of the fingers (a–d) and valleys (e–h) of the palm by conducting the method in Ref. [40]. Afterwards, we detect the maximum inscribed circle (T) of the palm to find the centre of the palm (see Figure 1.4c). Therefore, the location of the maximum inscribed circle in the enhanced image can be achieved (refer to Figure 1.4d). To acquire pixels from the background, four external tangent circles of T are located as B1, B2, B3, and B4 (see Figure 1.4c), which are on the vertical and horizontal directions. We define the radius of T as R; thus, the radiuses of B1, B2, B3, and B4 are defined as 0.5 × R, 0.5 × R, 0.5 × R, and 0.3 × R, correspondingly. Here, pixels in T are randomly selected as the positive data, and pixels in B1, B2, B3, and B4 are randomly selected as the negative data. Afterwards, the positive data and the negative data are put into the SVM to segment the palm from the background (see Figure 1.4e). 4. Contour Detection and ROI Extraction: Given the detected palm image (refer to Figure 1.4e), the Canny operator is utilised to achieve the boundary of the palm in the original image. Then, the boundaries named GAP1 and GAP2 between the forefinger and second finger and the fourth finger

(a)

(b)

(f)

(c)

(d)

(g)

(h)

FIGURE 1.4  Steps of hyperspectral palmprint ROI extraction [30].

(e)

8

AI and Deep Learning in Biometric Security

and little finger are obtained using the method in Ref. [41] (see Figure 1.4f), respectively. A line can then be drawn through one point in GAP1 and another point in GAP2, simultaneously. Then, we can define the two key points P1 and P2 when all points in GAP1 and GAP2 are below the line (see Figure 1.4f). Afterwards, a coordinate system is constructed based on P1 and P2 that the midpoint of line P1−P2 is defined as the origin O and a vertical line with P1−P2 passing O is defined as the x-axis (see Figure 1.4g). At last, a sub-image with a size of 128 × 128 in the palm centre is separated from the 3 image using the constructed coordinate system, where OC = P1 P2 (as seen 4 in Figure 1.4g and h). 1.3.1.2 Hyperspectral Dorsal Hand Vein ROI Extraction In the dorsal hand image, ROI indicates to the area that simply includes the vein part applied to extract feature. Dorsal hand vein images gathered through the acquisition device covers much redundant information such as a complicated background, the wrist, and the thumb. The unnecessary information can be eliminated by cropping the ROI from the collected image. The ROI not only maintains the vein structure with noise decreased but also reduces the computation cost, which can improve the recognition performance. The procedures of hyperspectral dorsal hand vein ROI extraction are presented in the following, which is adapted from our earlier study in Ref. [42] (refer to Figure 1.6): 1. Pinky Knuckle Point Detection: Based on a dorsal hand in the closed state (refer to Figure 1.6a), bulges at joints of the fingers and the boundary of a dorsal hand can be taken into consideration when locating the ROI. Here, the ROI can be extracted by locating one invariant point combined with a line of the profile of the dorsal hand. To this end, the template (refer to Figure 1.5) is constructed to search the point on a pinky knuckle. Based on a correlation operation between a template and a dorsal hand image, the maximal response (see the red point denoted in Figure 1.6f) can be found as the invariant point of the pinky knuckle. 2. Dorsal Hand Profile Location: The binarisation of a dorsal hand vein image was required for foreground segmenting from its background (refer to Figure 1.6b). Then, morphological opening and closing operations were applied to eliminate minor holes and remove tinny protrusions in the contour of the image (refer to Figure 1.6c). From the largest connected area (refer to Figure 1.6d), a profile of a dorsal hand (refer to Figure 1.6e) can be found by a boundary through single pixel-wise searching. 3. Key Line Determination: A circle was drawn with its centre at the point of the detected pinky knuckle, where the two crossing dots between the circle and the dorsal hand profile are located (refer to Figure 1.6f). A point was found concerning a lower area of a dorsal hand, which is connected with the pinky knuckle formed a closely horizontal line. Another point was searched regarding a higher reign of a dorsal hand, which is connected with the point on the pinky knuckle produced a closely vertical line. Here,

9

Multimodal Biometric Authentication System

FIGURE 1.5  The template to locate the pinky knuckle.

(a)

(b)

(c)

(d)

(f)

(g)

(h)

(i)

FIGURE 1.6  The steps of hyperspectral dorsal hand vein ROI extraction [42].

(e)

(j)

10

AI and Deep Learning in Biometric Security

we chose a horizontal line or vertical line in place of an edge of an ROI (refer to Figure 1.6g). 4. ROI Extraction: Finally, with the pinky knuckle point detected and one key line drawn, the other three edges of the ROI are determined (refer to Figure 1.6h). Due to the insufficient vein information in margin of a dorsal hand image, the ROI is moved a few pixels to the up and right to achieve rich vascular features (refer to Figure 1.6i). The experiments showed that this method is robust and adaptive at locating the ROI precisely for hyperspectral dorsal hand image (refer to Figure 1.6j).

1.3.2 Feature Extraction LBP is an effective and widely used texture feature descriptor [22] in biometric recognition. Not only does LBP obtain a better performance in many applications, it is also computationally simplistic [43]. Compared with LBP, LDP was proposed as a high-order texture encoding scheme for local patterns. Furthermore, LDP can extract the derivative direction variation information of each pixel in the image. The 2D-Gabor filter is sensitive to orientations, making it the most promising in the extraction of local palmprint and dorsal hand vein [2,5]. Otherwise, DCNN has obtained significant performances in image classification [44]. DCNN has powerful ability of abstract and impact feature representation by executing several nonlinear convolutional layers. Usually, abundant training data are necessary to train the parameters in the DCNN. Particularly, the derived characteristics from a certain layer can be utilised as the DCF for biometric authentication [10,44]. In this subsection, the classical feature extractors including LBP, LDP, 2D-Gabor, and DCNN are introduced as follows. Each will be utilised for the hyperspectral palmprint and dorsal hand vein ROIs (refer to Sections 1.4.2–1.4.4).

1. LBP: Texture has been proved an effective pattern in biometric recognition [9] due to its rich local characteristics. Given an ROI image, the key step to transform a pixel into the LBP code is to binarise its neighbouring eight pixels that the value of the centre pixel is chosen as the threshold. Afterwards, each pixel can be encoded as follows: d −1



LBPL , C =

∑S ( v − v ) × 2 (1.1) d

d

d =0



 0,  x > 0 S (x) =  (1.2)  1, x ≤ 0 where v is the value of the pixel at the location (L ,  C ) in the image, and vd is the value of the dth neighbour pixel. Finally, a LBP vector can be generated by using a histogram for all the encoded values. It is shown in Figure 1.7 that we can define the LBP descriptor with a variety of sizes (LBPd , r ), where d denotes the number of neighbour adjacent points and r denotes the radius.

11

Multimodal Biometric Authentication System

= 8,

=1

= 16,

=2

FIGURE 1.7  LBP neighbourhood sizes.

2. LDP: The LDP [9] is utilised to encode the local direction pattern. Given the ROI image I ( Z ) , we define its first-order derivatives on different orientations as I ∂′ ( Z ) , where ∂ = 0°, 45°, 90°, and 135°. Here, we assume that Z 0 is a point in I ( Z ) , and Zi (i = 1, …, 8) (see Figure 1.8) denotes the ith neighbour pixel. Therefore, the first-order derivatives of Z 0 is calculated as follows:

I 0′° = I ( Z 0 ) − I ( Z 4 ) (1.3)



I 45 ′ ° = I ( Z 0 ) − I ( Z3 ) (1.4)



I 90 ′ ° = I ( Z 0 ) − I ( Z 2 ) (1.5)



I135 ′ ° = I ( Z 0 ) − I ( Z1 ) (1.6) The second-order derivative of Z 0 on ∂ (∂ = 0°,  45°,  90°, and 135°) can be described as follows:



LDP∂2 ( Z 0 ) = { f ( I ∂′ ( Z 0 ) , I ∂′ ( Z1 )) ,  , f ( I ∂′ ( Z 0 ) , I ∂′ ( Z8 ))} (1.7)

FIGURE 1.8  Surrounding pixels around the centre point Z 0.

12

AI and Deep Learning in Biometric Security

where f (.,.) is an equation on binary transformation:

 0, if   I ∂′ ( Z 0 ) ⋅ I ∂′ ( Zi ) > 0  (1.8) f ( I ∂′ ( Z 0 ) ,  I ∂′ ( Zi )) =   1, if   I ∂′ ( Z 0 ) ⋅ I ∂′ ( Zi ) ≤ 0 At last, a 32-bit feature vector can be generated as follows on different orientations:



{

}

LDP 2 ( Z 0 ) = LDP∂2 ( Z 0 )| ∂ = 0°, 45°, 90° and 135° (1.9) 3. 2D-Gabor: Due to the fact that it has a good 2D spectral specificity property, the 2D-Gabor filter is frequently exploited in orientation feature extraction [2,5]. The 2D-Gabor is presented as follows:  x 2 + y2  1 exp −  exp {2πi( µ x cos ϕ + µ y sin ϕ )} 2 2πσ 2  2σ  (1.10)

G ( x ,  y,  ϕ ,  µ ,  σ ) =

where i = −1, µ presents the frequency of the sinusoidal wave, ϕ denotes the direction, and σ denotes the standard deviation. Usually, a 2D-Gabor filter bank contains a set of filters on n orientations with the same scale. The orientation ϕ j is obtained as follows:

ϕj =

π ( j − 1)  j = 1,  2,   …,  n. (1.11) n

Then, the convolution of the 2D-Gabor filter is conducted on the palmprint image to obtain a line response as follows:



(

rj = I *G (ϕ j )

)(

(1.12)

x , y )

where I denotes the image, G (ϕ j ) denotes the real part of the filter on ϕ j , “*” is the convolutional operator, r is the convolution result, and ( x ,  y ) denotes the position of a pixel in I. 4. DCF: DCNN usually includes a variety of components, such as pooling [11], convolution, ReLU [12], and Softmax-loss layer [23], as shown in Figure 1.9. LeCun [11] first utilised the LeNet on handwritten digit classification. Since then, DCNNs with the similar non-linear structure have been widely used [23]. There usually are thousands of parameters in different layers. Therefore, high impact and discriminative characteristics can be obtained after several convolutions with the trained parameters. Particularly, the Softmax-loss layer is used for classification as a classifier in DCNN. Here, we ignore the Softmax-loss layer and extract discriminative features as DCFs directly from the second FC layer (see Figure 1.9) of the DCNN.

13

Multimodal Biometric Authentication System

FIGURE 1.9  The architecture of DCNN for VGG-F [45].

1.3.3 Feature Fusion and Matching In the feature extraction phase (refer to Section 1.3.2), different features, including LBP, LDP, 2D-Gabor, and DCF, can be applied to each ROI image in the palmprint and dorsal hand vein cubes, respectively. If all images from the different bands are fused for recognition, it will be costly and time consuming. Consequently, we selected the optimal bands with respect to the types of features achieving the best recognition results on palmprint and dorsal hand vein, respectively. Let Fpalm = [ f1 ,  f2 , …,  fi , …,  fn ] ∈ R d × n and Pdhv = [ p1 ,  p2 , …,  ps , …,  pn ] ∈ R d × n denote the hyperspectral palmprint features and hyperspectral dorsal hand vein features, respectively, where fi is the feature vector for the ith band palmprint image, ps is the feature vector for the sth band dorsal hand vein ROI, d denotes the dimensionality of the feature, and n denotes the number of spectrums. Afterwards, the optimal features can be fused as follows:

W = O ( Fpalm ) ; O ( Pdhv ) (1.13)

where O (⋅) is the selection of the optimal feature from Fpalm or Pdhv, with the selected feature vector obtaining the highest recognition accuracy. Specifically, W is to be concatenated with the optimal O ( Fpalm ) and the optimal O ( Pdhv ). After feature fusion, we use the “Euclidean” distance for the final matching:

dist ( X , Y ) =

d

∑ ( x − y ) (1.14) 2

i

i

i =1

where X and Y are features extracted from two objects.

1.4 EXPERIMENTS First, we briefly introduce the collected hyperspectral palmprint and dorsal hand vein dataset. Then, the optimal band and pattern selection are performed on

14

AI and Deep Learning in Biometric Security

different modalities, respectively. Afterwards, multimodal identification and verification results are presented, correspondingly. At last, the time consumption of the proposed method is analysed.

1.4.1 Multimodal Hyperspectral Palmprint and Dorsal Hand Vein Dataset We constructed a hyperspectral palmprint and dorsal hand vein dataset captured from the same volunteers utilising the proposed hyperspectral imaging device (refer to Section 1.2). As mentioned in Section 1.2, the device can acquire hyperspectral images covering a spectrum range of 520–1,040 nm with 10 nm intervals, which means that the images on 53 different spectrums can be obtained. The dataset was acquired from 209 persons, and each volunteer was required to provide both left and right hands for imaging. This dataset contains two sessions which were acquired with intervals about 30 days. In each session, a volunteer was requested to capture both their left and right hands a total of five times. Therefore, this dataset totally includes 443,080 (209 subjects × 5 samples × 2 objects × 53 bands × 2 sessions × 2 modalities) images. Some original and ROI samples from one object are shown in Figures 1.10 and 1.11, respectively.

1.4.2 Optimal Pattern and Band Selection To obtain the best performance in recognition using multimodal features, we should select the best bands for palmprint and dorsal hand vein, respectively, in which the image contains rich and clear information and can derive the most discriminative features. For different feature patterns, including LBP, LDP, 2D-Gabor, and DCF, we aim to choose the optimal pattern and band for palmprint recognition and dorsal hand vein recognition, respectively. For every experiment, each algorithm was conducted 10 times. Finally, the mean accuracy of recognition rate was calculated as the performance evaluation:

ARR =

number of correctly classifised samples (1.15) total number of samples

(a)

(b)

FIGURE 1.10  Hyperspectral palmprint (a) and dorsal hand vein (b) samples.

15

Multimodal Biometric Authentication System

(a)

(b)

FIGURE 1.11  The hyperspectral palmprint and dorsal hand vein ROIs coming from the same individual. (a) denotes palmprint ROI samples and (b) denotes dorsal hand vein ROI samples. From left to right, top to down, the band increases from 520 to 1040 nm with 10 nm intervals.

In this work, the experiments were implemented using MATLAB 2015a on a CPU 3.40 with RAM 16.0 GB running Windows 10. When we extracted LBP and LDP features from the ROI image, each image was segmented into 16 non-overlapping sub-images with the same size of 32 × 32. Afterwards, the LBP or LDP features were extracted from each sub-image and further to be concatenated to one feature vector. As for 2D-Gabor, we defined a bank with five scales on eight directions. Otherwise, we applied VGG-F for DCF extraction with the DCF derived from the 19th layer of VGG-F. At last, the nearest neighbour (1-NN) was chosen for identification and verification. Figures 1.12 and 1.13 show the identification rates of different patterns on each band of the hyperspectral palmprint and dorsal hand vein cubes, respectively. From Figure 1.12a, one can see that LBP achieved the highest ARR (98.09%) on the 44th band corresponding to 950 nm. LDP (Figure 1.12b) obtained the highest ARR (94.74%) on the 41th band corresponding to 930 nm. 2D-Gabor (Figure 1.12c) achieved the highest ARR (76.08%) on the 37th band corresponding to 880 nm. DCF (Figure 1.12d) obtained the highest ARR (97.89%) on the 21th band corresponding to 730 nm. As for the dorsal hand vein results presented in Figure 1.13, one can see that LBP (Figure 1.13a) achieved the highest ARR on the 38th band corresponding to 880 nm with 92.20%. LDP (Figure 1.13b) achieved the highest ARR of 97.00% on the 52th band corresponding to 1030 nm. 2D-Gabor (Figure 1.13c) achieved the highest ARR on the 40th band corresponding to 900 nm with 88.20%. DCF (Figure 1.13d) obtained the highest ARR on the 26th band corresponding to 780 nm with 92.20%. Both Figures 1.12 and 1.13 show that different patterns have their own corresponding optimal bands. For hyperspectral palmprint identification, LBP can achieve the highest ARR on 950 nm with 98.09%. On the other hand, in hyperspectral dorsal hand vein identification, LDP can achieve the highest ARR on 1,030 nm with 97.00%.

16

AI and Deep Learning in Biometric Security

(a)

(b)

(c)

(d)

FIGURE 1.12  ARRs of different patterns for each band of the hyperspectral palmprint cube: (a) LBP, (b) LDP, (c) 2D-Gabor, and (d) DCF. When the ARR ≥ 0.9, the bar colour is black. With 0.7 ≤ ARR